AI Magazine Volume 9 Number 3 (1988) (© AAAI)

PENNSYLVANIA STATE CHESS CHAMPIONSHIP HITECH Becomes First Computer Senior Master

he Pennsylvania State Chess be 2405; so it has crossed the magical Hitech wins the Pennsylvania State Chess T Championship was held in 2400 boundary and is now a Senior Championship. Pittsburgh, on July 23-24. In a field of Master in the U.S. Chess Federation. 46 players, including 8 masters, It is the first computer to achieve this Carnegie-Mellon University's Hitech title. The above are subject to won the title outright with a score of verification by the USCF; however, in 4.5 - 0.5. The finale was very exciting the past our calculated ratings have as the number one seeded player, never differed from the USCF's by International Master Ed Formanek, more than 2 points. This advance also had a perfect 4-0 going into the final places Hitech among the top 150 play- round against the number two seed ers in the U.S. Hitech's performance Hitech, which was the only competi- rating was 2507. tor with 3.5 - 0.5. Hitech had to win In tournaments since the National to gain the title, whereas Formanek Open, Hitech had a mediocre result of could become Champion by drawing 3.5 - 2.5 at the Fredkin Masters Open this game. Hitech played beautifully, in Pittsburgh, May 28-30. Over the and the game culminated in a brilliant 4th of July holiday, Hitech competed endgame that Hitech handled perfect- in the World Open, Masters Section. ly. The game is appended below. The field was made up of 127 players, Hitech also finished first in last almost all of them Masters. Hitech year's State Championship, but after scored 5.5 - 2.5 and finished tied for the fact a number of issues arose relat- 10th. It lost its last round game, when ing to whether the computer could be winning would have meant tying for the State titlist. This year the rules 1st. specified that a computer could win It seems appropriate to indulge in a the title. So Hitech received the title, little retrospection at this point. The but was not eligible for the $400 first first program that was able to play prize money and was not supposed to chess at a level that could be mea- get the trophy either. However, in a sured on the National rating scale was fine gesture of sportsmanship, Profes- Greenblatt's MacHack-VI. It achieved sor Formanek of the Mathematics a rating of about 1440 (Class C) in the Department of Pennsylvania State 1966-1967 time period. In the period University, presented the trophy he 1970 - 1979, the Northwestern Uni- had won to Hitech since he indicated versity program CHESS4.X dominated Hitech clearly deserved it. Formanek the scene. Early versions of this pro- was one of four Masters that ended gram, by David Slate and Larry Atkin, tied for 2nd at 4-1. performed at approximately the 1600 This is the first time Hitech has level (Class B). When the CDC won a noncomputer tournament out- CYBER 176 came along, CHESS4.6's right; that is, without sharing 1st rating moved up dramatically to the place. Formanek, rated at 2485, is the vicinity of 1900 (Class A). In 1980, highest rated player Hitech has ever Ken Thompson and Joe Condon of beaten. It is the first time that Hitech Bell Labs built the special purpose has beaten an IM, and we believe only computer Belle. It immediately the second time that a machine has assumed the lead and eventually beaten an IM (Fidelity beat an IM). became a Master and set a new high We estimate Hitech's new rating to water mark for computers of 2203.

FALL 1988 85 Belle was dominant until 1983. resulting configuration. responsibility for the hardware and is Cray Blitz took over winning the Opponents are much tougher than designing some new units at the pre- computer tournaments at this point; they were before. About 1/3 of all sent time. however, it was never able to do much players exercise their option to refuse I have been responsible for doing with respect to the human rating to play computers. Of the remainder, the pattern knowledge (Berliner 1988) scale, something that we attribute to about 80% seem well prepared to play and most of the opening book, and the rather poor knowledge in that pro- Hitech. It is no longer unusual to find acting as moderator for the many fine gram. Beginning in 1985, Hitech start- people talking in the playing halls discussions that we have about how ed setting new records. In October, about what they believe to be to improve Hitech in the various areas 1985 Hitech eclipsed the old record Hitech's weaknesses. We have that need work. for computers by achieving a 2233 rat- encountered players who openly say Since August 1987, when Hitech ing. By December, the rating had been that they have collected and studied was reborn with new pattern recog- pushed to 2301. By July, 1986 Hitech's all of Hitech's games that they could nizers and properly working software, rating reached 2359. However, a great find, in case they had the opportunity it has played 48 US rated games and 4 deal of this was due to opponent's fail- to play it in the future. games in International matches. Its ing to treat Hitech with appropriate The first and third reasons above US performance rating over this peri- respect, and throwing away winning produce a slowing down of visible od is 2440. In two matches earlier this positions with careless play. The pre- progress. However, it allows Hitech's year, it beat M. Peretz of Israel 1.5 - sent high has been achieved some 22 developers to find those weaknesses 0.5, and drew M. Apicella of France 1- years after the 1966 MacHack-VI rat- that need to be fixed. Recently, Hitech 1. A summary of these results by cate- ing was established. It corresponds to has managed to produce ever better gory of player can be found below: a gain in rating of 965 points, or quality games and to continue to raise Expert or less (<2200): 15 - 0 approximately 45 points/year. This its rating as the result of improving its Master (2200-2399) 14.5 - 6.5 figure has also been arrived at by knowledge and searching capability Sr. Master (2400-2499) 3 - 1 other analysts. (but not speed). It is appropriate here Top 50 US 1 - 7 However, as the above shows, to give credit to those who have made Foreign 2.5 - 1.5 progress has not always been continu- this possible: Carl Ebeling (now at ous. If the present rate of progress University of Washington) built the Pennsylvania State Championship, continues, then we can expect a com- special purpose hardware (Ebeling Pittsburgh, July 24, 1988. puter as World Champion by the year 1985), and a good deal of software White : IM Formanek (2485) 1998. It is very difficult to predict relating to how to interface to the Black : Hitech (2385) exactly what is required to do this. hardware and see what it is doing for 1. e4, e5; New and very promising search algo- debugging purposes. 2. Nf3, Nc6; rithms are now appearing. Also, new Larry Slomer (now at PPG Indus- 3. Bb5, a6; hardware already has made possible tries) helped to build the hardware 4. Ba4, Nf6; speeds 6 times greater than Hitech's and maintained it after Carl graduat- 5. O-O, N:e4; in a single chip. However, it appears ed. 6. d4, b5; that the very best players will contin- Gordon Goetsch wrote most of the 7. Bb3, d5; ue to find weaknesses in computer system software that makes possible 8. N:e5 play until most of the chinks in the interfacing with the special purpose The most popular move here is 8. d:e5 knowledge armor are eliminated. hardware, and has recently revised the which keeps the Black N at c6 block- The fact that it has taken 30 software that supervises the search to ing its own pawn. I suspect our oppo- months to go from 2301 to the present make other searching disciplines pos- nent wanted to avoid well trodden 2405 can be attributed to several fac- sible. Gordon also maintains Hitech's paths that Hitech may have had much tors: The basic capability of a statistics. information on. machine is definitely related to its Murray Campbell has helped with 8., N:e5; speed and the depth to which it can the opening book, and has now imple- 9. d:e5, c6; search. Thus, as long as a machine mented the “Singular Search” 10. Nd2, cannot outsearch a class of human (Anatharam 1988) algorithm on Now Hitech is out of its book, and on opponents, it will likely remain inferi- Hitech. He has been my alter-ego its own. or to that class. At the moment when it comes to discussion of chess 10.,Nc5; Hitech can be outsearched by players ideas, and what can be done to allow 11. Nf3, Be7; in the class 2500+. Hitech to understand this or that. 12. c3, N:b3; As we learned more about the pat- Andy Palay has been involved with 13. a:b3, O-O; tern recognition process, we decided early design issues, and recently had Here Hitech had to judge whether to to rebuild our pattern recognizers. implemented the opening book in the play it safe with 13.-- Bg4 which This resulted in improvement in the form of a file system that allows restricts the N from any useful activi- play, but also in the loss of about one greater ease of access, and ty, or to allow what follows. It judged year of time, considering time to build modification. the sequel correctly. and install, and then to debug the Andy Gruss has now taken over

86 AI MAGAZINE 14. Nd4, Bd7; White could probably still hold the stop the two Black pawns, though this 15. b4, a5!; position by not moving any of his is not immediately obvious. White is trying to blockade the c- pawns. Usually, in such situations it 63. Nh3, Ke4; pawn by getting a N or B into c5, by is correct to keep your pawns on the 64. Ng1, Kf5; Nb3 and Be3 followed by Bc5. Hitech opposite color as your opponent's B. 65. Ne2, Kg4; understands such things, and adeptly However, here it weakens the square 66. Kg1, f3; foils all of White's attempts. g4 over which the Black K will even- 67. Nc3, Bc6; 16. R:a5, R:a5; tually infiltrate. I would have made 68. Kf2, K:h4; 17. b:a5, Q:a5; the same move under the circum- 69. Nd1, Kg4; 18. b4, Qa2; stances, but it is wrong. Black can 70. Ne3+, Kf4; 19. Be3, Qc4!; now win by force, and does. 71. Nf1, h4; Gaining time for the eventual c5 by 42., g5!; 72. Nh2, h3; attacking the pawn at c3. 43. g3, Bc8; If White did not have to move, he 20. Qa1, c5; 44. Nc6+, Kd6; could draw the game by just keeping 21. b:c5, B:c5; 45. Nd4, Bg4+; his pieces where they are and Black Black stands better. After White's 46. Kf2, Ke5; would not be able to advance either P. inferior next move, Hitech wins a P. 47. Nb5, Bd7; However, unfortunately for him, he is 22. h3?, Re8!; 48. Nd4, Ke4; in Zugzwang (the compulsion to 23. Nf3, 49. Nc2, Bb5; move). Hitech had already been pre- After 23. f4, f6 wins a P; and after 23. 50. Nd4, Ba6; dicting values of more than a R ahead Bf4,b4 also wins one. 51. Ne6, Bc8; for some time. 23.,B:e3; 52. Nd4, Bd7; 73. Nf1, Bb5; 24. f:e3, Qc5!; 53. Nc2, g:h4!; 74. Ng3, h2; 25. Qe1, Rc8; The critical decision. Now White has 75. Nh5+, Kg4; 26. Qf2, Be8; two weak pawns and the Black king 76. Ng3, Bf1!; 27. Nd4, Q:c3; will jockey around to win one of Not absolutely necessary, but nice. 28. N:b5!, Q:e5; them. The resulting ending with two The B cannot be taken, and if 77. Of course, not 28.-- B:b5; Q:f7+,Kh8; pawns ahead is then an easy win. Nh1, Bg2; Ng3,Kh3 and everything Qf8+ and mate next move. Despite 54. g:h4, Bg4; falls apart. Therefore: the fact that Black is a P ahead, this 55. Nd4, f5!; White Resigns position should be drawn. The reason A strange looking move that puts is that White's N is very strongly another P on the same color as his References placed inhibiting the movement of own B, which is a no-no in such situa- Ebeling, C., All the Right Moves — A VLSI Black's pieces, whereas Black's B is tions. However, as Hitech has fore- Architecture for Chess, MIT Press, 1986. unable to find effective targets. seen, this is only temporary, and the Anatharaman, T., Campbell, M., and Hsu, 29. Nd4, Rc3; move is actually the start of a pro- F., “Singular Extensions: Adding Selectivi- 30. Re1, Bd7; found winning maneuver. ty to Brute-Force Searching”, AAAI Spring 31. Qf3, Bc8; 56. Ne6, f4!; Symposium on Computer Game Playing, 32. Kf2, f6; The blockbusters start. If now 57. Palo Alto, 1988. 33. Ra1, Bb7; N:f4,d4!; Ng2 (N:h5,d:e+ preserving Berliner, H., and Ebeling, C. Forthcoming. 34. Rb1, Rc7; the last P, wins), d3; Ke1,Bf3 wins. “Pattern Knowledge and Search: The 35. Rb6, Kf7; 57. Nc5+, Kf5; SUPREM Architecture”, 36. Re6, Qg5; 58. Nd3, d4!!; Both sides have been jockeying Hitech had been playing very quickly, around without any clearly defined but I was unprepared for this shot aims. I was beginning to get con- which it made after only 16 seconds cerned that Hitech would mess things of reflection. If now 59. N:f4, Ke4 up; however, it is Formanek that wins as before. If 59. e:f, Ke4 winning cracks. His next move allows the the f-pawn and leaving Black with an swap of the major pieces, which easier win than in the game, because makes it possible for the Black K to the Black d-pawn is further removed participate in the fight. Then Black from the scene of action and thus has real winning chances. more dangerous. 37. Qg3??, Rc2+!; 59. e:d4, Ke4; 38. N:c2, Q:g3+; 60. Ne5, K:d4; 39. K:g3, K:e6; 61. Nf7, Be6; 40. Nd4+, Ke5; 62. Ng5, Bd5; 41. Kf3, h5!; Now the Black K goes over and 42. h4?, leisurely wins the White h-pawn. It is Post-game analysis revealed that impossible for the White K and N to

FALL 1988 87